Generating accessible multi-occupancy floor plans with fine-grained control using a diffusion model

Abstract

This paper presents a constrained transformer-based diffusion model for generating accessibility-aware and spatially coherent multi-occupancy building floor plans to support early-stage design. Existing approaches often produce low-resolution outputs and lack fine-grained control over room configurations. To address these issues, the proposed approach employs a latent transformer-based diffusion model to generate high-resolution floor plans with complex room shapes, diverse configurations, and numerous rooms while maintaining computational efficiency. It also integrates a flexible design constraint conditioning system, providing precise control over spatial and geometric constraints represented by room-level masks, bounding boxes, circles, and global-level boundaries. An iterative refinement workflow, guided by a rule-based accessibility checker, ensures compliance with selected accessibility requirements. Experimental results demonstrated significant improvements in floor plan quality and constraint control compared to the baseline. Case studies further showcased the approach's ability to support accessibility-aware design exploration while adhering to design constraints defined by the conditioning system.